signatures-hand-synergies

Check the write-up here. This project was done under Professor Kavita Ramanan and William Salkeld at Brown under the Spring 2023 UTRA Award.

File Organization

classify_stim.ipynb - Initial testing of classification algorithms

exploratory_data_analysis.ipynb - Initial plotting, analysis, etc. of data

main.ipynb - main results from writeup

deepmlp_classif.ipynb - similar results to main.ipynb but using MLP

Scripts:

normalizers.py - simple normalizers

paths.py - Path objects for different databases

Database Annotations

DB1

Associated paper

• Aquisition Setup

  • Kinematic Data
    • 22-sensor Cyberglove II representing 22 joint angles as 8-bit values at a resolution < 1 degree
    • 2-axis inclinometer fixed onto wrist to collect wrist orientation
    • 25Hz sampling frequency
  • Surface EMG
    • Double-differential MyoBock 13E200
    • 100Hz sampling frequency

• Stimulus: 52 movements divided into 4 main classes:

  • 12 movements of fingers (flexions and extensions)
  • 8 isometric, isotonic hand configurations/postures
  • 9 wrist movements (adduction/abduction, flexion/extension, pronation/supination)
  • 23 grasping and functional movements

• 27 subjects

• 10 repetitions of each class of movements

• 5 seconds of motion, 3 seconds of rest in-between

DB3

Associated paper

Collection of phantom limb electrical signals along forearm from hand-amputated subjects

• Aquisition Setup

  • Surface EMG
    • Double-differential MyoBock 13E200-50
    • 12 electrodes in total along different parts of forearm
    • 100Hz sampling frequency
    • Columns 1-8 are electrode signals around forearm
    • Columns 9 & 10 are signal along two activity spots of Flexor and Extensor Digitorum Superficialis
    • Columns 11 & 12 (partially -- "when available") are from electrodes on actiity spots of muscle Biceps Brachii and of the muscle Triceps Brachii

• 11 hand-amputated subjects

• 10 repetitions of each class of movements

• 5 seconds of motion, 3 seconds of rest in-between

• Contains 36 columns of data about (x, y, z) acceleration of 12 sEMG electrodes

• 2 columns of (roll, pitch) inclination values

• 6 columns of force values

• 2x6 columns of extremal force values (minimal and maximal force values for each sensor)

DB5

Associated paper

• Aquisition Setup

  • Kinematic Data
    • 22-sensor Cyberglove II representing 22 joint angles as 8-bit values at a resolution < 1 degree
    • 2-axis inclinometer fixed onto wrist to collect wrist orientation
    • 25Hz sampling frequency
  • Surface EMG
    • 2 Thalmic Myo bands, one tilted at 22.5 degrees above first
    • 16 electrodes
    • 200Hz sampling frequency
    • Columns 1-8 are the electrodes equally spaced around the forearm at the height of the radio humeral joint
    • Columns 9-16 represent the second Myo, tilted by 22.5 degrees clockwise.
    • 3 columns for accelerometer from first Myo
    • 200Hz sampling frequency

• Stimulus: 52 movements divided into 4 main classes:

  • 12 movements of fingers (flexions and extensions)
  • 8 isometric, isotonic hand configurations/postures
  • 9 wrist movements (adduction/abduction, flexion/extension, pronation/supination)
  • 23 grasping and functional movements

• 10 intact subjects

• 6 repetitions of each class of movements

• 5 seconds of motion, 3 seconds of rest in-between

• DB5 - For feature extraction and classification, "Repetitions 1, 3, 4 and 6 were used to train the classifiers, repetitions 2 and 5 were used for validating them. The classification was performed on all movements (rest included)" in DB5, according to this associated paper.

• DB5 sampling frequency is 200 Hz, so windowing into 200 sample-sized windows (with an overlap of 50%) involves 1 second of data in each window

DB7

Associated paper

Includes offline analysis of a real-time prosthetic hand control experiment with 12 subjects (11 intact, 1 amputee)

• Aquisition Setup

  • Does not collect any kinematic data
  • Surface EMG
    • Delsys Trigno IM Wireless EMG
    • 12 electrodes and 9-axes inertial measurement units (9 degrees of freedom: 3-axial accelerometer, gyroscope and magnetometer)
    • 100Hz sampling frequency

• Stimulus: 40 movements divided into 4 main classes:

  • 12 movements of fingers (flexions and extensions)
  • 8 isometric, isotonic hand configurations/postures
  • 9 wrist movements (adduction/abduction, flexion/extension, pronation/supination)
  • 23 grasping and functional movements

• 20 intact subjects, 2 amputees

• 6 repetitions of each class of movements

• 5 seconds of rest between movement trials

DB9

Associated paper

• Aquisition Setup

  • Kinematic Data
    • 22-sensor Cyberglove II representing 22 joint angles as 8-bit values at a resolution < 1 degree
    • 2-axis inclinometer fixed onto wrist to collect wrist orientation
    • 25Hz sampling frequency

• Stimulus: 40 movements divided into 4 main classes:

  • 8 isometric, isotonic hand configurations/postures
  • 9 wrist movements (adduction/abduction, flexion/extension, pronation/supination)
  • 23 grasping and functional movements

• 77 subjects

• 5 repetitions of each class of movements

• 5 seconds of motion, 3 seconds of rest in-between

• 22 columns of order of angles : name of the angles corresponding to variable “angles”

Zenodo Download

Evens DB Notes

• All have a stimulus, restimulus, repetition, rerepetition (re- is corrected for what acc happened, data can be ragged)
• EMG usually recorded with something attached to forearm
• If you wanted to combine databases, you would need to determine which exercises match across different databases since they all have a different ordering

DB2:

• 12 EMG columns, 2kHz
• 6 reps, 49 movements, 40 subjects (intact)
• Movements include hand positions (1-8), basic movements of wrists (9-17), grasps and functional movements (18-40), force patterns (41-49)
• 5 seconds + 3 seconds rest
• Glove is 22 dof version

• Sampling frequency is 2000 Hz

DB4:

Same paper as DB5, different instrument (Cometa vs. Double Myo)

• 12 EMG rather than 16 compared to DB5
• 3 exercises: (1) basic movements (2) wrist movements (3) grasping + functional movements
• 6 repetitions, 52 movements, 10 subjects (intact)
• 5 seconds + 3 seconds rest

• Sampling frequency is 2000 Hz

• eSMG dim is $10$

DB6

• "Repetability"
• unique thing is multiple days of acquisition - made participants do the movements twice a day for 5 days (larger dataset...)
• EMG - 16 dims (2 are empty though), 2 kHz
• 12 repetitions of 7 grasps only, 10 intact subjects
• 4 seconds + 4 seconds rest

DB8:

• 10 intacts, 2 amputees
• EMG - orig 1111 hz, then upsampled to 2khz, 16 dims
• Glove - 18 DoF
• 6-9 seconds + 3 seconds rest
• Each exercise is more like a "grip"/finger movement rather tahn an involved action
• Explicitly states that this database is meant for estimation/reconstruction of finger movement rather than movement/grip classification, since the data is meant to be slow finger movements and there is a lack of extended hold period). Though, that shouldn't affect using signatures, since its still tree-like equivalent to a properly timed movement...

DB10:

Newest but seemingly most involved (more details later, but I don't think it's that useful for the purposes of this project)

A lot of data that I probably can't store locally